Talk:Darkcanuck/RRServer

Fire away...

Just a suggestion for an additional check. I have never seen score a bot more than 8000 points, so this could be checked too. When examining the results that messed up the original roborumble rating beyond repair, I saw results of 20000 against 16000 (Thats what you get when running OneOnOne with MELEE=YES). For the time being I let my client running (unattended) for ABC's server, as I don't really have the time for bughunting. Your effort however seems promising. Good luck. -- GrubbmGait

• Thanks! That's a good check, will be combining that with the survival >=35 (also your suggestion I think) once I rearrange the error handling and failure output to the client. Then I'll look into ELO... --Darkcanuck
• Your checks have both been implemented. -- Darkcanuck

Looking very nice! I have a couple questions and thoughts I thought I'll mention. So what is this "Ideal" column in the results mean? One thought I had about ratings, is perhaps it would be best to make the APS fill missing pairings with Glicko-based estimates? I'm thinking that would give the best long term stability/accuracy once pairings are complete while having something a more meaningful before pairings are complete before the pairings are complete. --Rednaxela 01:18, 26 September 2008 (UTC)

Thanks! I've just posted a bit more about ratings here. The "Ideal" column is my attempt to reverse calculate a rating based on a bot's APS. I just inverted the Glicko formula for "E" (expected probability) to yield a rating given if given E (i.e. APS) and a competitor's rating and RD. For the latter two I used the defaults (1500 and 350) so theoretically if the APS represents the score vs an average bot (and there's a uniform distribution?) then the rating might converge to the "ideal" value. But I have no idea if it works, just wanted to see how close it might be. I'm not sure you could fill in the pairings using Glicko + APS -- the reason systems like Glicko exist is to get around the problem of incomplete pairings, so the Glicko rating should be enough in itself. If it's accurate, that is -- we'll see once the ratings catch up to the pairings already submitted... -- Darkcanuck 03:39, 26 September 2008 (UTC)

Ahh, I see. Thanks for the explanation. If the Glicko rating doesn't converge very very close to the "Ideal" then I'd say it alone might not be the best fit alone for Robocode due to how complete pairings are not hard to get. The reason I suggest using APS and filling missing pairings with Glicko-based percent estimates, is because my proposed method will be guaranteed to always converge to an exact APS ranking order when pairings are complete, and would quite surely be at least slightly better than APS when pairings are not complete. Perhaps I'm more picky than most, but I'd consider a hybrid necessary if "Glicko" doesn't in practice converge to "Ideal" to within an accuracy that preserves exact rankings with APS (which I think is very plain and simple the most fair when there are complete pairings). I suppose we'll see how accurately Glicko converges :) --Rednaxela 04:25, 26 September 2008 (UTC)

Be careful about the "ideal" convergence concept! Keep in mind that I made this value up and it doesn't really have a statistical basis of any sort. I was just curious what a naive reversal with a single data point might produce, in order to get an idea of what neighbourhood DrussGT's rating might be in, for example. I also wanted to get a sense whether I had programmed the formulas correctly. I wonder though, if we're abusing these rating systems by using %score instead of absolute win/lose values (1/0)? Would the Glicko rating converge more rapidly to match the APS scale if I had chosen win/loss? I'm very curious, but no so much as to interrupt the current rebuild, which may take longer than I thought. -- Darkcanuck 04:54, 26 September 2008 (UTC)

Well, I'm not talking about the convergence to that "Ideal" column. I'm talking about convergence of the relative rankings as opposed to specific rating numbers. If the rankings, don't converge to exactly the same order as APS, then I think there's issue enough to justify a hybrid that uses APS, with ELO or Glicko to estimate missing pairings. --Rednaxela 05:10, 26 September 2008 (UTC)

Gotcha. I suppose you could keep track of the rating (Elo or Glicko) and just use it to calculate expected scores for missing pairings. Then generate an estimated APS for full pairings. We'll have to see how well the ratings stabilize. I'm thinking I should have used Glicko-2 instead, since it includes a volatility rating to account for erratic (read problem bot) performance. -- Darkcanuck 06:22, 26 September 2008 (UTC)